method
active
method:mixed-clean-adversarial-training-protocol-0-5Mixed clean/adversarial training protocol (α=0.5)
Trains toy models for 150,000 steps with L2 ε=0.1‖x‖.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Adversarial Examples Are Not Bugs, They Are Superpositionintroducesmentions
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Unsatisfying reduction.
- Novelty claim.
- Intervention used to control robustness.
- Context-independent transformation.
- Reduces superposition but needs 80% L2 norm.
- Robustness result showing introspection stage's contribution for Llama model
- Robustness controls superposition.
- Adversarial examples activate more features than clean inputs across all ResNet18 layers.finding0.723Figure 6.